Using Machine Learning to Predict Form Processing Times : Applied to Swedish pay-as-you-earn tax returns

Sammanfattning: Forms are used in many situations. For example, they tend to be ubiquitous in communications between individuals and government agencies. Something which could potentially boost transparency and efficiency is accurate estimates of how long it will take for the receiver to process a given completed form. Unfortunately, such estimates are often not available. This thesis examines the problem of using machine learning to predict form processing times, applied to the context of Swedish pay-as-you-earn tax returns. More specifically, it compares a naive baseline model to several random forest models, some based on the more common batch learning principle, and others on online learning which is typically seen as more suitable for working with data streams and changing conditions. Despite the theoretical advantages of online learning, none of the models using that approach were able to consistently outperform the naive baseline model. Conversely, the two primarily evaluated batch learning models were successful in doing so, although the improvement over the baseline was small.